Persistent homology

Persistent homology converts changing shape into a record of features that appear and disappear across a filtration. A point cloud, image, network, or sampled geometric object is represented by a nested sequence of spaces, often built by adding simplices according to a scale parameter.¹ Homology summarizes each space algebraically: zero-dimensional classes represent connected components, one-dimensional classes represent loops, and two-dimensional classes represent enclosed voids. A class’s birth marks when it appears, while its death marks when it is filled in, merged, or otherwise becomes trivial. The interval between those events is its persistence. Long-lived classes may reflect broad structure, whereas short-lived classes are often associated with sampling variation or noise, although persistence alone does not establish scientific meaning.

Computational pipelines first build a filtration and then calculate the induced maps between homology groups. For a point cloud, common constructions include the Čech complex, Vietoris–Rips complex, and alpha complex; each makes different geometric or computational trade-offs.² Matrix-reduction algorithms transform boundary matrices into a pairing of birth and death events, producing a barcode or persistence diagram.¹³

The Vietoris–Rips complex is especially popular because it can be constructed from pairwise distances, but its simplex count can grow rapidly. Sparse approximations, coefficient choices, and truncation by homology dimension are therefore central practical decisions. Software implementations commonly exploit filtered-cell ordering and efficient reduction rather than constructing every possible simplex explicitly.

Persistence diagrams provide a compact summary whose geometry supports comparison between datasets. Each feature is represented by a point with birth and death coordinates, while the diagonal represents zero persistence; a barcode displays the same information as horizontal intervals. The bottleneck and Wasserstein distances compare such summaries, and stability theorems show that small perturbations of suitable input functions or metric data produce correspondingly small diagram changes.⁴⁵

Stability does not remove the need for modeling choices. The metric, filtration, sampling density, coefficient field, homology dimension, and treatment of essential classes all affect the result. Statistical procedures can assess persistence under resampling or noise models, while vectorizations such as persistence landscapes and persistence images make diagrams usable in conventional machine-learning workflows.

Persistent homology extends beyond the standard one-parameter barcode used for point clouds. Extended persistence records features through both ordinary and relative homology, allowing it to describe, for example, contours and voids in scalar fields more completely.⁶ Zigzag persistence permits maps that alternate between forward and backward inclusion, which is useful when observations are added and removed rather than monotonically accumulated. Multiparameter persistence studies several filtration parameters at once, but it generally lacks the complete interval decomposition available in the one-parameter setting.³

The method has been applied to sensor data, images, materials, dynamical systems, and biological data. In such settings it is best understood as a representation of multiscale structure, not as an automatic detector of a uniquely correct shape; domain knowledge remains necessary to connect a persistent feature with a mechanism.

Filtration

A nested sequence of spaces or complexes indexed by a parameter, with each stage containing the preceding stage.

Vietoris–Rips complex

A simplicial complex built from points whose pairwise distances satisfy a chosen scale condition.

Barcode

A collection of intervals showing the birth and death ranges of homological features.

Persistence diagram

A multiset of birth–death points representing the same persistence information as a barcode.

Bottleneck distance

A metric comparing persistence diagrams by minimizing the largest matching displacement, including matching to the diagonal.